PMC-VQA: 227,000 Medical Visual Question Answering Pairs

PMC-VQA contains 227,000 visual question-answering pairs associated with 149,000 medical images sourced from PubMed Central. Released by RadGenome and updated in July 2024, the collection includes a specialized version focused on noncompound images to facilitate cleaner model training. The dataset is organized into training and testing splits with a dedicated clean test set for benchmarking.

Use Cases

• Training multimodal transformers for medical reasoning using the VQA pairs in train.csv
• Benchmarking zero-shot medical image understanding using the test_clean.csv subset
• Developing image-to-text generation models specifically for noncompound medical figures in images2.zip

Strengths

• Large scale featuring 227,000 VQA pairs
• Includes 149,000 unique medical images
• Provides a test_clean.csv subset for high-reliability evaluation
• Version 2 specifically isolates noncompound images for improved clarity

Limitations

• Potential for label noise inherent in automated extraction from medical literature
• Geographic bias toward medical practices and data reported in US-based publications
• Complexity of compound images in version 1 may challenge standard VQA architectures

Provenance

Source

RadGenome (derived from PubMed Central)

Freshness

Last updated July 2, 2024.

Geography

United States